Author: Robert Hyatt
Date: 12:25:03 10/31/99
Go up one level in this thread
On October 31, 1999 at 14:20:23, blass uri wrote: >On October 31, 1999 at 13:57:54, Robert Hyatt wrote: > >>On October 31, 1999 at 11:53:38, blass uri wrote: >> >>>On October 31, 1999 at 10:12:06, Amir Ban wrote: >>> >>>>On October 30, 1999 at 17:52:02, Christophe Theron wrote: >>>> >>>>>On October 30, 1999 at 08:22:00, Enrique Irazoqui wrote: >>>>> >>>>>>I have played 2 matches at game/5 between Tiger 12.0 and Crafty 16.18 as an >>>>>>engine for Fritz. >>>>>> >>>>>>Crafty played on a PIII-500, 64MB hashtables, the Nalimov tablebases that come >>>>>>with Fritz and the General book of Fritz 5 built after games of 2500+ players. >>>>>> >>>>>>Tiger 12.0 played on a PII-300, 32MB hashtables and the small book of Tiger 11.7 >>>>>>with only 35000 positions. >>>>> >>>>>Oops... Not exactly. >>>>> >>>>>This book indeed comes from the first versions of Tiger 11.x but it contains >>>>>only 7682 moves. >>>>> >>>>>This is 35 times smaller than the current book provided with Tiger 12.0. >>>>> >>>>> >>>>> >>>>>> I used this book to compensate for Crafty not using >>>>>>its own. It was not uncommon to see Tiger out of book after 2, 3 or 4 moves. I >>>>>>don't think that the book gave Tiger any kind of advantage. >>>>>> >>>>>>In the first match, Tiger won 25-13, +19 -7 =12, scoring 65.7% >>>>> >>>>>Wow! What elo rating difference would that mean? >>>>> >>>>> >>>>> >>>>>>The second match was played under the same conditions, except that Tiger had PB >>>>>>off. In this second match, Tiger won 23-21, +16 -14 =14, scoring 52.2%. >>>>>> >>>>>>Going back to the discussion of a few weeks ago about PB on/off, these 2 matches >>>>>>seem to indicate that PB off is not more detrimental than what could be expected >>>>>>by just not using the usual 50% of the opponent's time. >>>>>> >>>>>>The delay in transmitting the moves through auto232 is almost 3 seconds/move for >>>>>>the dos driver and about 2/10 for the windows driver. Considering that the >>>>>>average in these matches is 79 moves/game, each game lasted 14 minutes instead >>>>>>of 10. Assuming that both programs guessed 50% of the opponent's moves, Tiger >>>>>>and Crafty used 9.5 minutes/game (5 + 4.5) each with PB on, while in the second >>>>>>match Tiger used 5 minutes/game. It is as if Tiger would have played the first >>>>>>match on a P300 and the second on a P150. All this mess (sorry) makes the >>>>>>results of both matches quite coherent. >>>>>> >>>>>>I tried all this PB on/off thing in a different way. Didzis plays with 2 >>>>>>programs on one machine and PB off. I replayed with 2 machines one of his games >>>>>>Tiger-CM6K and both programs played the same moves. >>>>>> >>>>>>So it seems that for some programs playing with PB off has no other effect than >>>>>>having less time to compute. >>>>> >>>>> >>>>>Also it seems that a crippled Tiger is still better than a full strength Crafty >>>>>(PII-300/small book against PIII-500). >>>>> >>>>>And it seems that a crippled crippled Tiger is still at least as strong as a >>>>>full strength Crafty (PII-300/PB off/small book against PIII-500). >>>>> >>>>>I find this interesting as some time ago Bob was laughing at me because I'm >>>>>still using a 386sx20 for some of my tests and algorithmic improvements. >>>>> >>>>>I would not be surprised if Chess Tiger 12.0 on PII-300 was able to stand Crafty >>>>>on a Quad-Xeon. After all that would only be a 4x speed advantage for Crafty. :) >>>>> >>>>> >>>>> >>>>> Christophe >>>> >>>> >>>>That's approximately the speed advantage in crafty vs. ban on ICC (that is, >>>>pre-noplaying and censoring): >>>> >>>>4 x 450 MHz at 5 + 3 inc vs. 2 x 350 MHz at 5 + (-1) (that's my setting because >>>>I'm manual). If the game lasts 60 moves that translates to: >>>> >>>> >>>> (1800 MHz * 8 min) / (700 MHz * 4 min) > 5 >>> >>>It is not exactly the case because crafty has not >5 times advantage in >>>pondering. >>> >>>Crafty can ponder 8 minutes when Junior can ponder the time that it does not >>>play (16-4 minutes=12 minutes) >>> >>>Uri >>> >> >>it is all bad math. The quad xeon runs (generally) a little over 3x faster >>than a single xeon would. There are many positions where it runs 4x faster, >>there are some where it is actually slower than a single processor. >> >>But in general, it is 3x faster. Amir is running on a dual 350, which is fairly >>close to 1/2 the xeon. The speed advantage is a little over 2, assuming his >>speedup for 2 is similar to mine. I have no idea how he would conclude anything >>greater. >> >>If he plays someone on a single cpu at 800 mhz, how much faster is that machine >>than his? I get 800/350 as a quantum estimate. But I would want benchmark >>numbers to be really happy, because the 800mhz box might have a memory bandwidth >>problem. Or a 133mhz bus advantage. > >It is more than 800/350 because the fact you play manually is a disadvabtage. >The fact that you can use the permanent brain does not fully compensate for the >time advantage. I disagree. I have played both manually _and_ automated. I didn't notice a lot of difference other than 'peace of mind'. Amir _has_ an automated interface, because he used to play on ICC automated. If manual is so bad, why doesn't he continue to use the auto interface? I have personally found that I could exert significant influence over crafty when I was manual. Forcing obvious moves, letting it search longer when things are difficult, overriding an occasional lemon move, etc. Running 100% automatic is _very_ difficult to do well. I assume Amir agrees since he went from automatic to manual, voluntarily. Automatic operation exposes you to a _lot_ that you don't get when manual. I think the difference is a toss-up so long as you avoid 0 inc games. As a quick sanity check, have amir play varguz. then compare my results to them (they are automatic) with my results to ban (which is manual). You might be surprised at how little automatic helps. > >simple math say if you assume ponder guessing of 50%,8 minutes/game(Junior is >using only 4 minutes for playing) >and every player is using 1/2 of the time that it does not use for playing and >correct in pondering. > >Junior is using 4 minutes+1/2*12 minutes=4+6=10 minutes. >crafty is using 8 minutes+1/2*8 minutes=12 minutes. > what is the "1/2*12? both programs have similar time limits. One is simply 2x faster than the other. Factoring in 'pondering' (which is random enough anyway) seems far less accurate than just comparing cpu speeds. I miss the concept of the 'handicap'. Since it is _constant_ for _all_ unequal hardware events. I always have 'more' time, since I am 2x faster. But overall it is _only_ 2x, not some fudged amount. >This is practically more than 12/10 time advantage because >it is better to use x seconds for every move than to use x/2 seconds when you >are wrong in pondering and 3x/2 when you are right in pondering. > >Uri
This page took 0.01 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.